OpenAI claims AI agents now complete nearly half of two-to-four-hour research tasks on their own — but says 'humans are still central to AI development'
OpenAI reports rising autonomous research performance, increased spending on coding agents, and persistent limitations that keep human researchers closely involved
In the span of seven months, starting from March 2026, AI agents have significantly increased their capacity to independently complete research assignments, covering a range of tasks that previously required several hours of human effort. According to OpenAI's internal assessment, these autonomous systems now handle 46% of assignments estimated to take two to four hours of human time, a jump from 23% back in March.
For longer assignments requiring four to eight hours of human labor, their success rate has more than doubled, from 13% to 35% during the same period.
However, the nature of the tasks influences the efficiency of AI agents. For instance, monitoring training runs is a task where agents excel, boasting a 79% success rate. Conversely, launching training runs and writing software for inference and serving systems prove more challenging, achieving only 25% success rates for the latter.
In terms of research categories, communication and analytical tasks see better results from autonomous agents, with success rates of 69% and 58% respectively, as opposed to software development which records a 43% success rate.
The growing role of AI agents in research also corresponds with increased spending on coding agents, with median daily expenditure reaching $2,400 in October, a 62% increase from September's $1,500. At the higher end, daily spending reached $23,000 for the top spenders, more than doubling September's figure of $11,000. Despite these advancements, OpenAI underscores that human researchers continue to play a crucial role, providing essential direction, supervision, and judgment throughout the development process.
Rewritten by urgent.news; may contain errors — check the original.